JP2007033898A - Speaker checking device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve collation precision without requiring a feature vector of a speech of another speaker even when a speech is distorted owing to background noise or under the influence of compression processing for the noise or when speaking contents of a speaker are different. <P>SOLUTION: A speech feature generation processing section 1 generates a feature vector by extracting a feature quantity from an input speech and also generates a feature representative vector by finding the center of gravity of a set of generated feature vectors. In collation mode, the first and the second speaker feature conversion processing sections 6 and 7 perform conversion processing for approximating code vectors and a feature vector set of a speaker to respective feature representative vectors and a threshold decision processing section 9 compares VQ distortion between vectors after the conversion processing with a threshold value to identify the speaker. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a speaker verification device that performs user authentication by extracting features of a user's voice, for example.

2. Description of the Related Art In recent years, various speaker verification devices have been proposed that perform user authentication by extracting features of a user's voice, for example, when entering a building or using equipment that requires a secure state.
For example, the input speech is sampled for each speaker, a feature vector is obtained from the feature amount, the feature vector is vector-quantized and registered in the codebook. Then, each time a user's voice is input, a feature vector is obtained based on the feature amount of the input voice, and the feature vector is vector-quantized and then collated with the vector registered in the codebook. There is known an apparatus for performing personal authentication (see, for example, Patent Document 1).

As another speaker verification device, a covariance matrix is calculated based on the feature vector of the speaker and the feature vector of the other speaker. An apparatus that statistically improves identification performance by converting feature vectors using the calculated covariance matrix has also been proposed (see, for example, Patent Document 2).
JP 7-248791 A Japanese Patent No. 3080388

  However, the apparatus described in Patent Document 1 simply collates the feature vector of the user's own voice registered in advance in the code book with the feature vector of the input speech. For this reason, if the speech is distorted by the influence of background noise or the utterance content of the speaker is different, the feature amount distribution of the other speaker is included in the feature amount distribution of the person speaker registered in the codebook. In some cases, the other speaker is misrecognized as the speaker, or the speaker is misrecognized as another speaker.

  On the other hand, the apparatus described in Patent Document 2 converts feature vectors based on a covariance matrix. For this reason, it is possible to raise collation precision compared with the apparatus described in Patent Document 1. However, in order to calculate the covariance matrix, it is necessary to acquire the feature vector of the voice of another speaker, and extra effort and cost are required for the processing.

  The present invention has been made paying attention to the above circumstances, and the purpose of the present invention is that even if the speech is distorted due to background noise or the effect of this suppression processing, and the utterance content of the speaker is different, It is an object of the present invention to provide a speaker verification apparatus that can improve verification accuracy without requiring a feature vector of speech.

  In order to achieve the above object, the present invention includes means for generating feature vectors from the input speech frames of speakers and generating feature representative vectors representing the plurality of generated feature vectors. For the speaker to be registered, the code vector and the feature representative vector generated by vector quantization of the generated feature vector are stored in the code book. On the other hand, when performing collation, a conversion process is performed on the first feature vector stored in the codebook to reduce the distance between the first feature vector and the first feature representative vector. In addition, a conversion process is performed to reduce the distance between the second feature vector and the second feature representative vector with respect to the second feature vector generated from the input speech of the verification target speaker. Then, the vector quantization distortion between the code vector after the conversion process and the second feature vector is calculated, the calculated vector quantization distortion is compared with a preset threshold value, and the comparison result is obtained. It is configured to output as a speaker verification result.

  Therefore, even if the input speech is distorted due to background noise or the effect of suppressing this, or the utterance content of the speaker is different, the registered speaker stored in the codebook is subjected to the above conversion processing on the feature vector. The distributions of the code vectors and the distribution of the feature vectors of the speaker to be collated are reduced. For this reason, the inclusion of the feature vector set of the verification target speaker with respect to the code vector of the registration target speaker is reduced. Therefore, the difference in vector quantization distortion between the code vector of the registration target speaker and the feature vector of the verification target speaker becomes significant, and this makes it possible to improve the verification accuracy.

  In short, in the present invention, for the first feature vector stored in the code book and the second feature vector generated from the input speech of the speaker to be verified, the first feature vector and its feature representative vector And a conversion process for reducing the distance between the second feature vector and the feature representative vector. Then, the vector quantization distortion between the code vector after the conversion process and the second feature vector is calculated, the calculated vector quantization distortion is compared with a preset threshold value, and the comparison result is obtained. This is output as a speaker verification result.

  Therefore, according to the present invention, even when the voice is distorted by background noise or the utterance contents of the speaker are different, the speaker capable of improving the collation accuracy without requiring the feature vector of the voice of another speaker. A verification device can be provided.

Embodiments of the present invention will be described below with reference to the drawings.
First, the principle of the present invention will be described.
Another person acceptance error that occurs in speaker verification using a Vector Quantization (VQ) model includes, for example, the feature vector of another person's speaker in the person's codebook as shown in FIG. Happens to be. In order to avoid this problem, intra-speaker distribution correction that reduces the distribution of the feature code of the person's codebook and the other person's speaker is effective. For example, as shown in FIG. 6, the distribution of the code vector and the feature vector may be made closer to the direction of each feature representative vector in each of the comparison target person's codebook and other person's feature vectors.

This intra-speaker distribution correction processing is realized by performing conversion processing for reducing the distance between the code vector and the feature vector, respectively, between the vector and the feature representative vector. For example, the feature vector before conversion processing is v _i , the feature vector after conversion processing is v _i ′, the feature representative vector is v _g , the conversion matrix is A, the first weight matrix is B, and the second weight matrix is C When the number of vectors is N, the conversion equation is expressed by the following equation (1).

In this equation (1), the transformation matrix A is represented by a DxD diagonal matrix whose diagonal component diag {A} is 0.0 or more, for example. The first and second weight matrices B and C are, for example, DxD unit matrices. D represents the number of vector dimensions.

  The feature representative vector can be expressed, for example, as a centroid of a code vector and a feature vector set. In this way, by making the center of gravity of the code vector and the feature vector set as feature representative vectors, the influence of the background noise or this suppression processing, the variation of the code vector and the feature vector due to the difference in the utterance content, can be absorbed, It becomes possible to make it depend only on the speaker.

When each diagonal elements of the diagonal matrix A described above and a _d, the range of values of a _d in order to reduce the distribution of feature vectors is 0.0-1.0, wherein the a _d This is called the reduction rate. For example, when a _d = 1.0 is set for all dimensions, the reduction ratio a _d is in a non-conversion state. On the other hand, if a _d = 0.0 is set for all dimensions, only the feature representative vector is set as the speaker feature vector.
FIGS. 7 and 8 show examples of speaker dependence of feature representative vectors. FIG. 7 shows a feature representative vector of speaker X, and FIG. 8 shows a feature representative vector of speaker Y, respectively. FIG. 9 shows the feature representative vectors of speaker X and speaker Y superimposed on each other.

As is clear from the figure, the dimension in which the feature representative vector varies greatly among the speakers is a dimension that more significantly represents the difference in the features between the speakers. On the other hand, the dimension in which the variation of the feature representative vector is small is a dimension in which the feature difference between the speakers is small. Therefore, the reduction ratio a _d for large variation dimension set to a value close to 0.0, whereas for small variations dimension to set the reduction ratio a _d to a value close to 1.0. Calculation formula of the reduction ratio a _d is expressed for example as the following equation (2).

Here, σ _d is a standard deviation (d = 1, 2,... D) of feature representative vectors between speakers for each dimension, and indicates in which dimension speaker characteristics tend to appear in the feature representative vectors. The larger the value, the more likely it is that the speaker will appear. p is an adjustment parameter for the dimensionality average reduction ratio a _d to B, not possible to specify values directly, as represented by the following equation.

In the equation, B is a dimension average value of the reduction rate a _d and is a parameter for reducing the inclusion of the speaker feature vector. The greater the value of B, the less the effect of reducing inclusion. In contrast, the smaller the value, the greater the effect of reducing the inclusion. However, if the value is too small, the increase in VQ distortion due to the variation of the speaker's feature representative vector becomes significant, and a person's permission error is likely to occur. The parameter B is preferably set to a value around 0.5, for example, and it is desirable to avoid excessive correction by making the value too small.

q is a parameter for controlling how much the difference of the standard deviation σ _d for each dimension is emphasized, and the degree of emphasis becomes maximum when q = 0.0. On the other hand, when q = B, there is no difference for each dimension, and the reduction ratio is a constant dimension value = B (in this case, p = 0.0).
The standard deviation σ _d can be set to an appropriate value by previously analyzing and calculating feature vectors of a plurality of speakers in a plurality of environments. Here, if the calculated value is obtained by analyzing speech under a sufficiently large number of speakers and noisy environment, it can be regarded as the universal nature of the feature vector extracted from the speech signal. it can. That is, this value only needs to be calculated in advance, and the value need not be calculated afterwards according to the use environment. Further, since a cepstrum coefficient such as an LPC cepstrum is used for the feature vector, the difference between the analysis environment and the voice input system including the microphone in the real environment appears as a linear difference of the feature vector. Therefore, the difference in the voice input system is absorbed when calculating the standard deviation σ _d of the feature representative vector between speakers.

Next, an embodiment of a speaker verification apparatus according to the present invention will be described.
FIG. 1 is a block diagram showing the functional configuration. This speaker verification device includes a speech feature generation processing unit 1. The speech feature generation processing unit 1 generates a feature vector from the input speaker's speech, obtains the center of gravity of the generated feature vector set, and sets this center of gravity as a feature representative vector.

FIG. 2 is a block diagram showing a functional configuration of the voice feature generation processing unit 1. The speech feature generation processing unit 1 includes a preprocessing unit 11, an LPC coefficient calculation unit 12, an LPCC generation unit 13, and a speaker feature representative vector generation unit 14.
The preprocessing unit 11 performs an analog-digital (A / D) conversion and noise suppression processing on the input speech signal, sets a speech analysis section, and sets a speech waveform in the analysis section for a certain time and Cut out by the analysis window at a fixed shift period, and generate and hold an audio frame.

  The LPC coefficient calculation unit 12 extracts a feature amount related to the personality information included in the audio signal by linear predictive coding (LPC) from each audio frame formed by the preprocessing unit 11. The LPCC generation unit 13 generates a feature vector (LPC cepstrum, first-order or higher parameters not including the 0th power term) based on the feature amount extracted by the LPC coefficient calculation unit 12. The speaker feature representative vector generation unit 14 calculates the center of gravity of the feature vector set generated by the LPCC generation unit 13. The calculated center of gravity is used as a feature representative vector.

  The speaker verification apparatus also includes a vector quantization unit 2 and a speaker-specific codebook database 3 as functions for speaker registration processing. The vector quantization unit 2 takes in the feature vector set generated by the speech feature generation processing unit 1 and its feature representative vector in a state where the speaker registration mode is selected. Then, vector quantization is performed on the feature vector, and a code vector is output. For example, as shown in FIG. 3, the speaker-specific codebook database 3 includes a code vector generated by the vector quantization unit 2 and a feature representative vector generated by the speech feature generation processing unit 1. The information is stored in association with the speaker-specific name ID input by the name input means.

  Further, the speaker verification device includes a conversion coefficient / threshold control unit 4, a conversion coefficient / threshold database 5, and a first speaker feature conversion processing unit as functions necessary for the speaker verification process. 6, a second speaker feature conversion processing unit 7, a VQ distortion calculation unit 8, a threshold determination processing unit 9, and an end determination unit 10.

  The conversion coefficient / threshold control unit 4 uses the conversion coefficient / threshold database 5 to obtain a speaker feature conversion coefficient for reducing the other-person acceptance error rate and a threshold for VQ distortion determination corresponding to this value. Read. The speaker feature conversion coefficient indicates the reduction rate a. The reduction ratio “a” is a value prepared in advance in consideration of the variation of the feature representative vector for each dimension between speakers in accordance with the above-described equation (2). The threshold is set so that the person acceptance rate is equal to the stranger rejection rate. For example, as shown in FIG. 4, the conversion coefficient / threshold database 5 stores the above-mentioned calculated and set speaker feature conversion coefficients and determination thresholds.

  The first speaker feature conversion processing unit 6 reads out a code vector and a feature representative vector for each speaker from the speaker-specific codebook database 3 when the speaker verification mode is selected. Then, a conversion process is performed to bring the code vector closer to the feature representative vector by using the conversion formula shown by the expression (1) and the conversion coefficient given from the conversion coefficient / threshold control unit 4.

  The second speaker feature conversion processing unit 7 takes in the speaker feature vector set and the feature representative vector generated by the speech feature generation processing unit 1 when the speaker verification mode is selected. Then, a conversion process for bringing the feature vector set of the speaker closer to the feature representative vector using the conversion formula shown by the expression (1) and the conversion coefficient given from the conversion coefficient / threshold control unit 4 I do.

The VQ distortion calculation unit 8 takes in the converted code vector and the converted speaker feature vector from the first and second speaker feature conversion processing units 6 and 7, respectively. Then, VQ distortion is calculated between the converted code vector and speaker feature vector.
The threshold determination processing unit 9 compares the VQ distortion calculated by the VQ distortion calculation unit 8 with a threshold given from the conversion coefficient / threshold control unit 4, and a flag signal indicating the comparison result Is output.

  The end determination unit 10 determines whether or not the speaker to be verified is the person himself / herself based on the flag signal output from the threshold value determination processing unit 9. When the person is determined to be the person, the conversion coefficient and threshold value are updated by giving a conversion coefficient change control signal to the conversion coefficient / threshold value control unit 4. The conversion processing by the person feature conversion processing units 6 and 7 and the VQ distortion determination processing by the threshold determination processing unit 9 are repeatedly executed. And if the said repetition process is performed the preset number of times, the comprehensive determination result will be output as collation determination information.

Next, the operation of the apparatus configured as described above will be described.
First, prior to collation, the voice characteristics of the speaker to be collated, that is, the speaker of the person himself, are registered. That is, when the speaker himself inputs his / her voice from the microphone, this voice is converted into a voice frame by the voice feature generation processing unit 1, and then LPC analysis is performed for each frame. Is extracted. A set of feature vectors is generated based on the feature amount. Further, the center vector of the generated feature vector set becomes the feature representative vector of the feature vector set.

The generated feature vector set is vector quantized by the vector quantization unit 2 and is stored in the speaker-specific codebook database 3 together with the feature representative vector in association with the speaker-specific name ID. Similarly, a speech feature vector set and its feature representative vector are also generated for other verification target speakers, and the code vector and feature representative vector of the feature vector set are stored in the speaker-specific codebook database 3.
The conversion coefficient / threshold database 5 stores a set of conversion coefficients and threshold values obtained by conducting a pre-experiment in advance.

  Now, when the registration process to each of the databases 3 and 5 is completed as described above, the speaker voice collating process is subsequently executed as follows. That is, when the speech of the speaker to be collated is input, the speech feature generation processing unit 1 generates a set of feature vectors of the input speech and a feature representative vector thereof. The feature vector set and the feature representative vector are Input to the speaker feature conversion processing unit 7. The second speaker feature conversion processing unit 7 uses the conversion formula previously given by the expression (1) and the conversion coefficient given from the conversion coefficient / threshold control unit 4 to describe the speaker characteristics. A conversion process for bringing the vector set closer to the feature representative vector is performed.

  At the same time, the first speaker feature conversion processing unit 6 reads a code vector and its feature representative vector for each speaker from the codebook database 3 for each speaker. Then, using the conversion equation shown by the equation (1) and the conversion coefficient given from the conversion coefficient / threshold control unit 4, a conversion process is performed to bring the code vector close to the feature representative vector. .

  The VQ distortion calculation unit 8 calculates a VQ distortion between the converted speaker feature vector set and the code vector stored in the speaker-specific codebook database 3, and the calculated VQ distortion is calculated. The threshold value determination processing unit 9 compares the threshold value. The determination result is output to the end determination unit 10 as a flag signal.

  The end determination unit 10 determines whether or not the speaker to be verified is the person himself / herself based on the flag signal. When it is determined that the user is the person himself / herself, a conversion coefficient change control signal is given from the end determination section 10 to the conversion coefficient / threshold control section 4. As a result, the conversion coefficient and threshold value control unit 4 updates the conversion coefficient and threshold value. This update is performed by shifting the conversion coefficient and the threshold value by a certain amount in a direction to reduce the error rate of accepting others. Each of the first and second speaker feature conversion processing units 6 and 7 performs conversion processing of the code vector and the speaker feature vector set using the updated conversion coefficient. Then, the VQ distortion between the converted code vector and the speaker feature vector set is calculated by the VQ distortion calculation unit 8, and the calculated VQ distortion is compared with the threshold value in the threshold determination processing unit 9. The

  The end determination unit 10 determines whether or not the speaker to be verified is the person himself / herself based on the flag signal representing the comparison result. A coefficient change control signal is provided. Then, the first and second speaker feature conversion processing units 6 and 7 further perform conversion processing of the code vector and speaker feature vector set based on the updated conversion coefficient. The VQ distortion between vectors is compared with the threshold value in the threshold value determination processing unit 9.

  Thereafter, each time the speaker to be verified is determined to be the person in question, the conversion coefficient and the threshold value are sequentially updated in a direction to reduce the error rate for accepting the other person. In addition, a series of collating processes from the conversion process of the code vector and the speaker feature vector set to the VQ distortion comparison process are repeatedly performed. If the number of repetitions reaches one or more times set in advance, and the final comparison result at that time is the person himself, a determination result indicating that the speaker is the person is output. On the other hand, if it is determined that the speaker is another person during the repetition, a determination result indicating that the speaker is another person is output. In this case, it may be determined that the collation is impossible, and the user may be prompted to restart the collation process from the beginning.

  As described above, in this embodiment, the speech feature generation processing unit 1 extracts feature amounts from input speech to generate feature vectors, and obtains a feature representative vector by obtaining the center of gravity of the generated feature vector set. It is trying to generate. Then, in the collation mode, the code vector and speaker feature vector set based on the conversion coefficient given from the conversion coefficient / threshold control unit 4 by the first and second speaker feature conversion processing units 6 and 7. Is converted to each feature representative vector, and the VQ distortion between the vectors after the conversion processing is compared with the threshold value by the threshold value determination processing unit 9, thereby determining the identity of the speaker. .

  Therefore, even if the input speech is distorted due to background noise or the effect of this suppression processing, and the utterance contents of the speaker are different, the distribution of the code vectors stored in the speaker-specific codebook database 3 and the verification target speaker's Each feature vector distribution is reduced. For this reason, the inclusion of the feature vector set of the speaker to be verified with respect to the code vector of the registered speaker is reduced, and thus the difference in vector quantization distortion between the code vector of the registered speaker and the feature vector of the speaker to be verified is reduced. Becomes conspicuous, and this makes it possible to improve collation accuracy.

  That is, the apparatus of this embodiment has high robustness in a noisy environment. As shown in FIG. 11, the feature vector in a noisy environment tends to approach the distribution of noise-only feature vectors as compared to a clean speech feature vector. This can also be imagined from the fact that the greater the noise, the more it is buried in the noise and only the noise can be heard. As a result, there arises a problem that speaker verification performance deteriorates in a noisy environment. In order to solve this problem, noise suppression processing is generally performed on noise-added speech. However, when the noise suppression process is performed, excessive suppression occurs as illustrated in FIG. 11, and the distribution of the feature vector is widened, which causes a side effect of causing the inclusion of the distribution as illustrated in FIG.

  On the other hand, in the embodiment of the present invention, since the conversion process for reducing the distribution of the feature vector is performed on the feature vector, the inclusion relationship of the distribution of the feature vector due to the side effects as described above is enhanced even when the noise suppression process is performed. As a result, performance degradation due to background noise can be reduced.

Furthermore, in this embodiment, when the speech feature generation processing unit 1 generates the feature vector set and feature representative vector of the input speech, only the voiced sound is targeted. For this reason, the unvoiced sound that is easily affected by the background noise can be excluded in advance from the feature extraction target, thereby reducing the utterance dependency.
Furthermore, the speech feature generation processing unit 1 calculates the centroid of the feature vector set, and uses this centroid as the feature representative vector. In this way, the deviation of each feature vector due to the utterance bias is canceled by taking the average, and as a result, the variation of the feature representative vector can be reduced.

  In addition, in the embodiment, when the speaker is determined to be the person as a result of the comparison between the VQ distortion and the threshold value, the conversion coefficient and the threshold value are sequentially updated in a direction in which the error acceptance rate of others is reduced, A series of collating processes from a code vector and speaker feature vector set conversion process to a VQ distortion comparison process are repeatedly executed a plurality of times. For this reason, both the person rejection rate and the other person acceptance error rate can be kept low.

FIG. 10 shows an example of the effect of the apparatus according to this embodiment. FIG. 10 is a line graph showing the average misrecognition rate when there is no feature value conversion, when the reduction rate a is constant for each vector dimension, and when the reduction rate a is different for each dimension. Is.
As is apparent from FIG. 10, by using the conversion process according to this embodiment, the inclusion relation of the feature vector distribution of the other person's speaker with respect to the person's codebook is reduced, and the average misrecognition compared to the case without conversion. The rate can be improved.

  The present invention is not limited to the above embodiment. For example, in the embodiment, the feature vector and the representative vector are generated only for the voiced sound of the input voice. However, the present invention is not limited to this, and feature vectors and representative vectors may be generated only for vowels. That is, a normal voice signal always includes a vowel, and this vowel is an element that significantly represents individuality. Therefore, using the time continuity of vowels, those corresponding to vowels are extracted from the feature vectors, and the average of the code vectors obtained by quantizing the feature vectors for each vowel is used as the feature representative vector. In this way, it is possible to reduce the deviation of the feature representative vector due to the utterance bias.

In addition to the LPC cepstrum, the feature vector can be implemented by applying the feature vector generated from MFCC and spectrum analysis and the matching model to not only VQ but also other matching models such as GMM.
In addition, the number of repetitions of a series of matching processes from the conversion process of the code vector and speaker feature vector set to the VQ distortion comparison process, the configuration of the conversion formula, the value of the parameter, and the like do not depart from the gist of the present invention. Various modifications can be made within the range.

  In short, the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. In addition, various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, you may combine suitably the component covering different embodiment.

The functional block diagram which shows one Embodiment of the speaker collation apparatus concerning this invention. The functional block diagram which shows the structure of the audio | voice feature production | generation process part of the speaker collation apparatus shown in FIG. The figure which shows an example of the information element memorize | stored in the speaker codebook database of the speaker collation apparatus shown in FIG. The figure which shows an example of the information element memorize | stored in the conversion coefficient and threshold value database of the speaker collation apparatus shown in FIG. The figure which shows an example of the feature vector space before conversion correction | amendment. The figure which shows an example of the feature vector space after conversion correction | amendment. The figure which shows the change of the feature representative vector with respect to the vector dimension of the speaker X. The figure which shows the change of the feature representative vector with respect to the vector dimension of the speaker Y. The figure which shows the dispersion | variation between speakers of the feature representative vector with respect to a vector dimension. The figure which shows the change of the misrecognition rate when there is no conversion correction, the vector dimension is constant, and the vector dimension is variably set to three patterns. The figure which shows an example of the relationship between feature vector distribution and noise distribution.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Speech feature production | generation process part, 2 ... Vector quantization part, 3 ... Codebook database classified by speaker, 4 ... Conversion coefficient and threshold value control part, 5 ... Conversion coefficient and threshold value database, 6, 7 ... Story Person feature conversion processing unit, 8 ... VQ distortion calculation unit, 9 ... threshold value determination processing unit, 10 ... end determination unit, 11 ... preprocessing unit, 12 ... LPC coefficient calculation unit, 13 ... LPCC generation unit, 14 ... talk Person feature representative vector generation unit.

Claims (8)

A feature vector generating means for time-dividing a speaker's input speech into a plurality of frames and generating a feature vector from each of the divided frames;
Feature representative vector generation means for generating a feature representative vector representing a plurality of feature vectors generated by the feature vector generation means;
For a speaker to be registered, a code vector obtained by vector quantization of the first feature vector generated by the feature vector generating unit, and a first feature representative vector generated by the feature representative vector generating unit A codebook to remember,
For the speaker to be verified, between the second feature vector generated by the feature vector generation unit and the second feature vector generated by the feature representative vector generation unit. Means for performing a conversion process to reduce the distance of
Means for performing a conversion process for reducing the distance between the code vector and the first feature representative vector with respect to the code vector stored in the code book;
Means for calculating a vector quantization distortion between the code vector after the conversion process and the second feature vector after the conversion process;
A speaker verification apparatus comprising: a determination unit that compares the calculated vector quantization distortion with a preset threshold value and outputs the comparison result as a speaker verification result.   The means for performing the conversion processing obtains a matrix obtained by multiplying a transformation matrix having a feature vector transformation function by a difference between a feature vector and a weight feature representative vector obtained by multiplying the first weight matrix, and for this matrix, The speaker verification apparatus according to claim 1, wherein the conversion operation is performed by adding a load feature representative vector multiplied by the second weight matrix.   The means for performing the conversion processing calculates each diagonal component of the diagonal matrix from the average value of the diagonal component and the variation of each dimension of the feature representative vector when the conversion matrix is a diagonal matrix. The speaker verification apparatus according to claim 2, wherein: The feature vector generation means includes
Means for extracting a voiced sound frame from a plurality of frames of input speech;
The speaker verification apparatus according to claim 1, further comprising means for generating a feature vector from the extracted voiced sound frame. The feature representative vector generation means includes:
Means for extracting a feature vector consisting only of a voiced sound frame from a plurality of feature vectors generated by the feature vector generation means;
The speaker verification apparatus according to claim 1, further comprising means for calculating a feature representative vector from a feature vector consisting only of the extracted voiced sound. The feature representative vector generation means includes:
Means for extracting a feature vector corresponding to a vowel of the input speech from among a plurality of feature vectors generated by the feature vector generation means;
The speaker verification apparatus according to claim 1, further comprising means for calculating a feature representative vector from a feature vector corresponding to the extracted vowel. The feature representative vector generation means includes:
Means for vector-quantizing each of the plurality of feature vectors generated by the feature vector generator;
The speaker verification apparatus according to claim 1, further comprising means for calculating centroids of the plurality of vector-quantized code vectors and using the calculated centroids as feature representative vectors. Means for variably setting both the conversion coefficient used for the conversion process and the threshold value;
Each time both the transform coefficient and the threshold value are variably set, the transform process for the code vector and the second feature vector, and the calculation of the vector quantization distortion between the code vector after the transform process and the second feature vector Means for repeating the processing and processing for comparing the calculated vector quantization distortion with a threshold;
The speaker verification apparatus according to claim 1, further comprising means for obtaining a verification result based on a plurality of determination results obtained by the repetitive processing.

Images